Opportunities and computational challenges in large-scale whole-genome sequencing data analysis

Genomic selection has been used in animal breeding for c. 15 yr and continues to be an important tool in predicting genetic merit in livestock populations. The dairy cattle industry was the first to adopt genomic selection, initially based on some 50K single-nucleotide polymorphism (SNP) arrays for thousands of animals. Later advances in genome-scanning technologies have enabled inexpensive genotyping and sequencing, leading to wider adoption, and constantly increasing amounts of genomic data, both as to the number of genotyped animals and variants genotyped per animal. Full sequence data are expected to supersede SNP chips in the coming years. We review the methods and computational approaches used with sequence data and the impact of the methods and model assumptions on genomic prediction accuracy. The modeling, development, and applicability of these methods to sequence data are discussed, as well as the computational resources required. Sequence data should, in principle, provide full information on genetic variability, which should lead to higher prediction accuracy. In practice, there is limited evidence of additional benefit from using sequence data over medium- or high-density SNP panels. This is particularly true for small effective population sizes (Ne) such as cattle populations, where animals within a breed have many common ancestors and thus longer chromosome segments with high linkage disequilibrium accurately trackable with a relatively small number of markers. A population with a small Ne has long haplotype blocks, from 1 to 5 Mb, making it hard to identify causal variants within blocks. However, in major cattle breeds, a medium-density SNP panel is sufficient to tag the blocks themselves, and prediction with large datasets is highly accurate. Clearly, sequence data should not be used directly for genomic prediction, but for identifying putative causal variants to improve the accuracy and stability of subsequent predictions. We show that the best strategy to deal with any large data with high SNP densities is to use only a subset of (important) markers and determine the most appropriate model for exploiting the preselected variants in the genomic evaluation. Novel prediction methods that subset trait-specific informative markers could offer the advantage of using sequence data by potentially linking individuals through underlying functional variants rather than simply through shared haplotype blocks inherited from ancestors. Further research is required to clarify this aspect.